Skip to main content

GuideArtificial intelligence policy and regulationRevision 1Updated

ExplainerAI GovernanceEuropean Union· 10 min read· in Artificial Intelligence

The Complete Guide to Artificial Intelligence Policy and Regulation

Global artificial intelligence governance relies on a patchwork of risk-based legislation, sectoral guidelines, and international treaties. This guide examines how different jurisdictions classify algorithmic risk, assign liability, and attempt to regulate autonomous systems.

By Nicolas Laurent

Share this story

Listen

In short

  • The European Union regulates AI through a horizontal, risk-based framework that imposes strict pre-market testing on high-risk systems.
  • The United States relies on a decentralized, sectoral approach where existing federal agencies enforce rules within their specific domains.
  • International organizations use soft law, such as the OECD AI Principles and UN resolutions, to establish global norms without direct enforcement.

Artificial intelligence policy represents the attempt by governments to govern software that learns, adapts, and makes autonomous decisions. Unlike traditional software regulation, which focuses on data privacy and cybersecurity, AI governance addresses the probabilistic nature of machine learning models.

The core challenge for lawmakers is the pacing problem. Regulatory frameworks take years to draft, debate, and implement, while artificial intelligence capabilities advance in months. This temporal mismatch forces governments to regulate the application of the technology rather than the underlying mathematics.

Consequently, global AI policy has fractured into distinct regional philosophies. Our analysis of the dominant regulatory frameworks reveals that jurisdictions are currently testing whether algorithmic systems are best governed as consumer products, as specialized services, or as entirely new legal entities requiring novel oversight mechanisms.[14]

The Global Regulatory Triad

Three dominant regulatory models currently shape the global artificial intelligence landscape. The European Union treats artificial intelligence as a product safety issue, the United States relies on decentralized sectoral enforcement, and China focuses on algorithm registries and state security.

The European approach relies on horizontal legislation that applies across all industries. The EU AI Act, published in the Official Journal in July 2024, establishes a uniform legal framework for the development, market placement, and use of AI systems across the 27 member states.[9]

In contrast, the United States utilizes a sectoral approach. Rather than creating a single comprehensive AI law, the US directs existing federal agencies to apply their current statutory authority to algorithmic systems within their specific domains, such as healthcare, finance, and transportation.[7]

"My Administration places the highest urgency on governing the development and use of AI safely and responsibly," states the November 2023 US Executive Order on Safe, Secure, and Trustworthy Artificial Intelligence. The directive mandates that agencies establish specific guidelines for their respective sectors.[7]

China's regulatory framework targets specific algorithmic applications rather than the underlying technology. The Cyberspace Administration of China requires developers of recommendation algorithms and generative AI services to register their models with the state and undergo security assessments before public release.

These three models create a complex compliance environment for multinational technology companies. A developer deploying a global medical diagnostic model must navigate pre-market conformity assessments in Europe, post-market agency scrutiny in the United States, and state security reviews in China.

The Anatomy of Risk-Based Regulation

The most widely adopted mechanism for AI governance is the risk-based framework. This approach categorizes artificial intelligence systems based on the potential severity of harm they pose to individuals, society, or critical infrastructure, applying stricter rules to higher-risk applications.

The European Union's framework divides systems into four tiers: unacceptable risk, high risk, limited risk, and minimal risk. Systems deemed an unacceptable risk, such as government social scoring or real-time biometric categorization in public spaces, are entirely prohibited within the bloc.[9]

High-risk systems face the most rigorous compliance burden. These include algorithms used in critical infrastructure, educational admissions, employment sorting, and law enforcement. Developers must conduct fundamental rights impact assessments, maintain detailed technical documentation, and ensure human oversight before deployment.[9]

Limited risk systems, such as chatbots and deepfakes, primarily face transparency obligations. The law requires developers to clearly label synthetic content so users know they are interacting with a machine. Minimal risk systems, which comprise the vast majority of AI applications, remain largely unregulated.[9]

The European Union's risk-based framework imposes the strictest requirements on high-risk and unacceptable applications.

Canada's Artificial Intelligence and Data Act adopts a similar risk-based philosophy but leaves the specific definition of high-impact systems to subsequent regulatory drafting. The March 2023 companion document emphasizes that the government will define high-impact categories based on the severity and probability of harm.[11]

Penalties for violating these risk-based frameworks are severe. Under the EU AI Act, companies deploying prohibited systems can face fines of up to €35 million or seven percent of their total worldwide annual turnover, whichever is higher.[9]

This penalty structure mirrors the enforcement mechanism of the General Data Protection Regulation. By tying fines to global revenue, regulators ensure that even the largest multinational technology companies feel the financial impact of non-compliance.

The EU AI Act ties maximum penalties to global revenue, mirroring the enforcement structure of the GDPR.

Sectoral and Decentralized Governance

The United States relies on a decentralized model, arguing that the risks of artificial intelligence are highly contextual. An algorithm predicting equipment failure in a manufacturing plant poses fundamentally different risks than an algorithm predicting criminal recidivism in a courtroom.

The National Institute of Standards and Technology provides the foundational architecture for this approach. The January 2023 AI Risk Management Framework offers voluntary guidelines for organizations to map, measure, and manage algorithmic risks throughout the system lifecycle.[1]

"AI systems are inherently socio-technical in nature, meaning they are influenced by societal dynamics and human behavior," the NIST framework states. It emphasizes that risk cannot be eliminated entirely, but must be managed through continuous monitoring and organizational accountability.[1]

Federal agencies use these voluntary standards to inform their mandatory enforcement actions. The Federal Trade Commission polices deceptive AI marketing claims, the Securities and Exchange Commission monitors algorithmic trading, and the Food and Drug Administration regulates AI-enabled medical devices.

This sectoral approach allows for rapid regulatory adaptation. When a new algorithmic harm emerges in the financial sector, the relevant agency can issue targeted guidance without waiting for a divided legislature to pass comprehensive horizontal legislation.

However, decentralized governance creates regulatory gaps. If an AI application falls outside the clear jurisdiction of an existing federal agency, it may operate with minimal oversight. This fragmentation has prompted several US states to draft their own comprehensive AI legislation.

The United States relies on existing federal agencies to enforce AI regulations within their specific sectors.

International Consensus and Soft Law

Because artificial intelligence development crosses borders instantly, international organizations rely heavily on soft law—non-binding principles and declarations—to establish global norms. These instruments lack direct enforcement mechanisms but shape the drafting of binding national legislation.

The Organisation for Economic Co-operation and Development established the first major intergovernmental standard. The OECD AI Principles emphasize inclusive growth, human-centered values, transparency, robustness, and accountability in the deployment of algorithmic systems.[2]

The United Nations Educational, Scientific and Cultural Organization previously laid the groundwork for this consensus. In 2021, UNESCO adopted the Recommendation on the Ethics of Artificial Intelligence, providing a comprehensive normative framework adopted by 193 member states to guide the ethical development of algorithmic systems.[6]

The United Nations General Assembly reinforced these norms by adopting a landmark resolution on artificial intelligence. The resolution urges member states to bridge the artificial intelligence divide and ensure that algorithmic systems promote sustainable development and human rights.[5]

"The same rights that people have offline must also be protected online, including throughout the life cycle of artificial intelligence systems," the UN resolution declares. It represents the first time the global body reached a consensus on the governance of the technology.[5]

The Council of Europe took a step further by drafting the Framework Convention on Artificial Intelligence. This treaty creates binding legal obligations for signatory states to ensure that AI systems respect human rights, democracy, and the rule of law.[3]

In November 2023, 28 nations signed the Bletchley Declaration during the inaugural AI Safety Summit in the United Kingdom. The declaration acknowledges the potential for serious, even catastrophic, harm arising from the most highly capable frontier AI models.[8]

In November 2023, 28 nations signed the Bletchley Declaration, acknowledging the potential risks of frontier AI models.

Corporate Governance and Standards

Statutory law moves too slowly to govern the daily engineering decisions of artificial intelligence laboratories. Consequently, the industry relies heavily on technical standards and internal corporate governance mechanisms to operationalize abstract policy goals.

The International Organization for Standardization published ISO/IEC 42001 in December 2023. This standard provides a certifiable management system for artificial intelligence, allowing organizations to demonstrate that they have implemented rigorous risk management and impact assessment protocols.[13]

Internal governance often relies on red-teaming, a practice adapted from cybersecurity. Specialized teams deliberately attempt to bypass a model's safety filters to identify vulnerabilities, biases, or dangerous capabilities before the system is released to the public.

The United Kingdom's approach heavily emphasizes this technical evaluation. The March 2023 policy paper outlines a pro-innovation framework that relies on regulators issuing context-specific guidance rather than imposing rigid statutory rules on developers.[4]

"We will not put the AI principles on a statutory footing initially," the UK government paper states. "New rigid and onerous legislative requirements on businesses could hold back AI innovation and reduce our ability to respond quickly and in a proportionate way."[4]

"We will not put the AI principles on a statutory footing initially," the UK government paper states.

Singapore's Model AI Governance Framework provides another blueprint for corporate accountability. Published in January 2020, the framework translates ethical principles into practical implementation steps, focusing on internal governance structures, human involvement in decision-making, and operations management.[10]

Frontier AI and Compute Thresholds

The most intense regulatory debate centers on frontier AI—highly capable foundation models that could possess dangerous capabilities. Because these models are general-purpose, regulators struggle to predict how they will be used or what specific harms they might cause.

To identify these models, policymakers increasingly rely on compute thresholds. This metric measures the total amount of computational power, expressed in floating-point operations, used to train a model. Larger training runs generally produce more capable and potentially riskier systems.

The US Executive Order establishes a reporting threshold of ten to the twenty-sixth power floating-point operations. Companies training models that exceed this massive computational threshold must notify the federal government and share the results of their safety tests.[7]

The European Union utilizes a similar mechanism for general-purpose AI models. Systems trained using computational power exceeding ten to the twenty-fifth power floating-point operations are presumed to carry systemic risks and face stringent auditing and cybersecurity requirements.[9]

Regulators use computational power thresholds to identify frontier models that require mandatory safety reporting.

Relying on compute thresholds provides regulators with a clear, measurable metric. However, algorithmic efficiency improves rapidly. A model trained next year might achieve frontier-level capabilities using significantly less computational power, potentially slipping beneath the established regulatory thresholds.

Liability, Copyright, and the Courts

As artificial intelligence systems generate text, images, and code, they challenge the foundational principles of intellectual property law. Courts globally are currently determining whether synthetic media can be copyrighted and whether training models on copyrighted data constitutes infringement.

The United States Copyright Office maintains that human authorship is a fundamental requirement for copyright protection. Consequently, images generated entirely by a text-to-image model cannot be copyrighted, placing synthetic media squarely in the public domain.

The United Kingdom offers a rare statutory exception. A 1988 provision in British copyright law grants protection to computer-generated works that lack a human author, assigning the copyright to the person who made the arrangements necessary for the creation of the work.

Liability presents an equally complex challenge. When an autonomous system causes physical or financial harm, the legal system must determine whether the developer, the deployer, or the user bears responsibility.

Traditional product liability law holds manufacturers strictly liable for design defects. However, many artificial intelligence systems are deployed as software-as-a-service, a category that courts have historically shielded from strict product liability claims.

The Stanford Artificial Intelligence Index Report 2024 highlights the rapid escalation of these legal disputes. The report notes that AI-related regulations in the United States grew significantly, alongside a sharp increase in federal lawsuits concerning algorithmic copyright infringement and liability.[12]

AI-related litigation and regulatory actions have surged globally as courts grapple with copyright and liability disputes.

The Trajectory of AI Policy

Artificial intelligence policy is transitioning from a period of abstract ethical declarations into an era of concrete statutory enforcement. The coming years will test whether risk-based frameworks can effectively govern autonomous systems without stifling technological innovation.

Regulatory divergence remains the greatest challenge for the global technology industry. As the European Union enforces strict compliance mandates, the United States relies on market forces, and China tightens state control, the dream of a unified global AI governance framework appears increasingly remote.

Ultimately, the success of these policies will depend on regulatory capacity. Governments must hire auditors, data scientists, and machine learning engineers capable of inspecting complex neural networks, ensuring that statutory mandates translate into verifiable algorithmic safety.

Comparison

Structural comparison of the three dominant global AI regulatory models.
FrameworkPrimary MechanismEnforcement StyleCore Philosophy
European UnionHorizontal risk tiersPre-market conformity assessmentsFundamental rights protection
United StatesSectoral guidelinesPost-market agency enforcementMarket-driven innovation
ChinaAlgorithm registriesState security reviewsInformation control and state alignment

How we did this

Method
Cross-mapping the classification criteria for 'high-risk' AI systems across the EU AI Act, the NIST AI RMF, and Canada's AIDA to identify regulatory overlap and divergence.
What we found
While all three frameworks target biometric identification and critical infrastructure, only the EU explicitly categorizes educational admissions and employment sorting algorithms as inherently high-risk by default, whereas the US and Canadian frameworks treat these as context-dependent risks requiring organizational self-assessment rather than mandatory external audits.
What we worked from
Limits of this analysis
Regulatory definitions are subject to ongoing judicial interpretation and agency rulemaking, meaning practical enforcement may align more closely than statutory text suggests.

Key terms

Compute threshold
A regulatory metric based on the total computational power used to train an artificial intelligence model, used to identify highly capable systems.
Red-teaming
A safety evaluation practice where specialized testers deliberately attempt to bypass an artificial intelligence model's safeguards to uncover vulnerabilities.
Conformity assessment
A formal evaluation process required by the European Union to demonstrate that a high-risk artificial intelligence system meets all statutory safety requirements.
Sectoral regulation
A governance approach where existing agencies apply rules to artificial intelligence within their specific industries, rather than creating a single overarching law.

Reader questions

Does the EU AI Act apply to companies based outside of Europe?

Yes. The legislation has extraterritorial reach, meaning it applies to any developer or deployer whose artificial intelligence system's outputs are used within the European Union, regardless of where the company is headquartered.

What is the difference between hard law and soft law in AI policy?

Hard law consists of binding statutes and regulations with financial penalties for non-compliance, like the EU AI Act. Soft law includes voluntary guidelines, declarations, and principles, such as the OECD AI Principles, which shape norms but lack direct enforcement.

How do regulators define a frontier AI model?

Regulators typically define frontier models based on the computational power used to train them. The US and EU use thresholds of 10^26 and 10^25 floating-point operations, respectively, to identify models that may possess dangerous or unpredictable capabilities.

Can an AI system hold a copyright for the content it generates?

In most jurisdictions, including the United States, copyright requires human authorship, placing AI-generated content in the public domain. However, the United Kingdom provides a specific statutory exception that grants copyright to the human who arranged for the computer-generated work.

What changed

  1. · Revision 1Published.
Precautionary Regulators 35%Innovation Advocates 30%International Consensus Builders 20%Corporate Governance Specialists 15%
Precautionary Regulators
Argue that AI systems require strict pre-market testing and fundamental rights assessments to prevent systemic societal harm.
Innovation Advocates
Favor sectoral, post-market enforcement and voluntary standards to avoid stifling technological growth with rigid statutory rules.
International Consensus Builders
Focus on establishing global norms and bridging the AI divide through soft law and multilateral treaties.
Corporate Governance Specialists
Emphasize internal risk management, red-teaming, and certifiable management systems to operationalize abstract policy goals.

Perspectives this story doesn't cover

  • Open-source developers advocating for regulatory exemptions
  • Civil rights organizations focused on algorithmic bias in marginalized communities

Sources

Source coverage

14 outlets

4 viewpoints surfaced

Precautionary Regulators 35%Innovation Advocates 30%International Consensus Builders 20%Corporate Governance Specialists 15%
  1. [1]National Institute of Standards and TechnologyInnovation Advocates

    Artificial Intelligence Risk Management Framework (AI RMF 1.0)

    Read on National Institute of Standards and Technology →
  2. [2]OECD Legal InstrumentsInternational Consensus Builders

    Recommendation of the Council on Artificial Intelligence

    Read on OECD Legal Instruments →
  3. [3]Council of EuropePrecautionary Regulators

    Council of Europe Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law (CETS No. 225)

    Read on Council of Europe →
  4. [4]GOV.UKInnovation Advocates

    AI regulation: a pro-innovation approach

    Read on GOV.UK →
  5. [5]United Nations Digital LibraryInternational Consensus Builders

    Seizing the opportunities of safe, secure and trustworthy artificial intelligence systems for sustainable development : resolution / adopted by the General Assembly

    Read on United Nations Digital Library →
  6. [6]UNESCOInternational Consensus Builders

    Recommendation on the Ethics of Artificial Intelligence

    Read on UNESCO →
  7. [7]Federal Register

    Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence

    Read on Federal Register →
  8. [8]GOV.UKInnovation Advocates

    The Bletchley Declaration by Countries Attending the AI Safety Summit, 1-2 November 2023

    Read on GOV.UK →
  9. [9]EUR-LexPrecautionary Regulators

    Regulation (EU) 2024/1689 of the European Parliament and of the Council laying down harmonised rules on artificial intelligence

    Read on EUR-Lex →
  10. [10]Personal Data Protection Commission SingaporeCorporate Governance Specialists

    Model AI Governance Framework

    Read on Personal Data Protection Commission Singapore →
  11. [11]Innovation, Science and Economic Development Canada

    The Artificial Intelligence and Data Act (AIDA) – Companion document

    Read on Innovation, Science and Economic Development Canada →
  12. [12]Stanford University

    Artificial Intelligence Index Report 2024

  13. [13]International Organization for StandardizationCorporate Governance Specialists

    ISO/IEC 42001:2023 Information technology — Artificial intelligence — Management system

    Read on International Organization for Standardization →
  14. [14]Factlen Editorial Team

    Synthesis by Factlen editorial team

    Read on Factlen Editorial Team →
Listen

Comments

Stay informed

Every angle. Every day.

Get Artificial Intelligence stories with full source coverage and perspective breakdowns, free every day.